Advanced Techniques for Mobile Robotics Gaussian...

Post on 07-Aug-2019

214 views 0 download

transcript

Wolfram Burgard, Cyrill Stachniss,

Kai Arras, Maren Bennewitz

Gaussian Processes in Robotics

Advanced Techniques for Mobile Robotics

Overview

§  Regression problem

§  Gaussian process models

§  Learning GPs

§  Applications

§  Summary

The Regression Problem §  Given n observed points

§  Assuming the dependency

§  How to predict new points

The Regression Problem §  Given n observed points

The Regression Problem §  Solution 1: Parametric models

§  Linear

§ Quadratic

§ Higher order polynomials

§ …

§  Learning: optimizing the parameters

The Regression Problem §  Solution 1: Parametric models

The Regression Problem §  Solution 1: Parametric models

The Regression Problem §  Solution 2: Non-parametric models

§  Radial Basis functions

§ Histograms, Splines, Support Vector Machines …

§  Learning: finding the structure of the model and optimize its parameters

The Regression Problem §  Given n observed points

The Regression Problem §  Solution 3: Express

directly in terms of the data points

§  Idea: Any finite set of values sampled from has a joint Gaussian distribution with a covariance matrix given by

Gaussian Process Models §  Then, the n+1 dimensional vector

which includes the new target to be predicted , comes from an n+1 dimensional Gaussian

§  The predictive distribution for the new target is a 1-dimensional Gaussian

§  Given the n observed points §  Squared exponential covariance

function

§  with §  and a noise level

Gaussian Process Model

The Regression Problem §  Given n observed points

Gaussian Process Models §  GP model

Learning GPs §  The squared exponential

covariance function:

§  Easy to interpret parameters

amplitude

index/input distance

characteristic lengthscale

noise level

Learning GPs §  Example: low noise

Learning GPs §  Example: medium noise

Learning GPs §  Example: high noise

Learning GPs §  Example: small lengthscale

Learning GPs §  Example: large lengthscale

Learning GPs §  Covariance function specifies the prior

prior posterior

Gaussian Process Models §  Recall, the n+1 dimensional vector

comes from an n+1 dimensional normal distribution

§  The predictive distribution for the new target is a 1-dimensional Gaussian.

§  Why?

The Gaussian Distribution §  Recall the 2-dimensional joint Gaussian:

§  The conditionals and the marginals are also Gaussians Figure taken from

Carl E. Rasmussen: NIPS 2006 Tutorial

The Gaussian Distribution §  Simple bivariate example:

The Gaussian Distribution §  Simple bivariate example:

joint

marginal conditional

The Gaussian Distribution § Marginalization:

The Gaussian Distribution §  The conditional:

The Gaussian Distribution §  Slightly more complicated in the general

case:

§  The conditionals and the marginals are also Gaussians

Figure taken from Carl E. Rasmussen: NIPS 2006 Tutorial

The Gaussian Distribution §  Conditioning the joint Gaussian in general

§  In case of zero mean:

Gaussian Process Models §  Recall the GP assumption

Gaussian Process Models §  Noise-free mean and variance of the

predictive distribution have the form

§  with

Gaussian Process Models §  Mean and variance of the predictive

distribution then lead to

§  with

Learning GPs §  Learning a Gaussian process means

§  choosing a covariance function §  finding its parameters and the noise level

§  What is the objective?

Learning GPs §  The hyperparameters

can be found by maximizing the likelihood of the training data e.g., using gradient methods

Learning GPs §  Objective: high data likelihood

§  Due to the Gaussian assumption, GPs have Occam’s razor built in

data fit complexity penalty

const.

Occam‘s Razor §  Use the simplest explanation that is

needed to describe the data

§  Data-fit favors overfitting §  Complexity penalty favors simplicity

too long just right too short

Advanced Topics / Extensions §  Classification/non-Gaussian noise §  Sparse GPs: Approximations for large

data sets §  Heteroscedastic GPs: Modeling non-

constant noise §  Nonstationary GPs: Modeling varying

smoothness (lengthscales) § Mixtures of GPs §  Uncertain inputs §  …

Further Reading

Applications in Robotics § Monocular range sensing §  Terrain modeling §  Learning sensor models §  Learning to control a blimp §  Localization in cellular networks §  Time-series forecasting §  …

Applications in Robotics §  Monocular range sensing §  Terrain modeling §  Learning sensor models §  Learning to control a blimp §  Localization in cellular networks §  Time-series forecasting §  …

Monocular Range Sensing

§  Can we learn range from single, monocular camera images?

Training Setup §  Mobile robot + laser range finder §  Omni-directional monocular camera

Training Setup DFKI Saarbrücken University of Freiburg

Learning Range from Vision §  Associate (polar) pixel columns with ranges

Extract features

Associate with ranges

Pre-processing §  Warp images into a panoramic view

§  120 pixels per column

§  Transform to HSV -> 420 dimensions

Visual Features §  Two types of features

1. No human engineering: Principle components analysis (PCA) on raw input

2. Use of domain specific knowledge: Edge features that shall correspond to floor boundaries

Experiments

Typical 180° scan

Online Prediction

Mapping Results Laser-based Vision-based

Saarbrücken:

Freiburg:

GP-based Terrain Modeling §  3D terrain models are important in

many tasks in outdoor robotics

Terrain Modeling §  Given: observations of the terrain

surface §  Task: Learn a predictive model §  Classic Approach: Elevation grid maps

GP-Based Approach §  Generalize the grid-based model to

fully continuous spaces by viewing the problem as function regression

§  Requirements § Probabilistic formulation to handle

uncertainty § Ability to adapt to the spatial

structures

Covariance Function §  Standard covariance function have

limited flexibility to adapt to the local spatial structure

strong smoothing

medium smoothing

little smoothing

Covariance Function § What is optimal in this case?

Local Kernel Adaptation §  Adapt kernels based on the terrain

gradients §  Covariance is adjusted according to the

change in terrain elevation in the local neighborhood

local average

elevation gradient

Adapting to Local Structures

Ground truth

Stationary GP Non-stationary GP

Adapting to Local Structure §  Model to deal with slowly changing

characteristics and strong discontinuities

Experiments

standard

adaptive

Experiments

Observation (with white noise σ=0.3)

Kernels Predicted Map Local errors

Experiments – Stone Block

Experiments – Stone Block

Ground Truth Observations

Experiments – Stone Block

Prediction

Experiments – Slope

Uncertainties:

Observations & model:

Summary §  GPs are a flexible and practical

approach to Bayesian regression §  Prior knowledge is encoded in a

human understandable way §  Learned models can be interpreted §  Efficiency mainly depends on the

number of training points §  Real-world problem sizes require

approximations/sparsity/…